Just finished D&D session with my group without much automations - just dndbeyond and primitive maps. Game went much better, broken software didn’t take away attention. I guess in TTRPGs no automation > bad automation.

Rookframe progress: scenes, a miniature library with physics, and—most importantly—plugins can now define UI in TSX and use the VTT API to roll dice or change world state. That resolves my biggest PoC unknown. Next: multiplayer and a Mörk Borg plugin I can test with my players =)

Got a bug in graphic settings menu that got me stuck on the black screen. Now I see why we need „confirm setting or they will be rolled back in 10 seconds“ 😅

Speedrunning my expiring codex resets - 4 worktrees, each compiles rust and stress tests the app as part of e2e. Makes laptop fans go crazy. And falling asleep with those sounds really brings up „Serial Experiments Lain“ vibe

Rookframe progress this week: - fixed a leak between Bevy and Bun: RAM under stress went from ~1 GB to ~500 MB - dice rolling is much nicer now - first UI iteration, static dummy for now Also dropped the reinvented performance-measurement system and 24K lines of code. Great success =)

Was working on UI-rendering problem: plugins in Bun environment should be able to drive UI in Rust app over the TCP. Nor Unity nor Unreal docs clicked for this problem. Then I found Shopify’s remote-rendering work. Same constraint, same architecture shopify.engineering/remote-rende...

Remote rendering: Shopify’s take on extensible UI - Shopify

A deep dive into the latest generation of technology that allows developers to extend Shopify’s UI.

shopify.engineering

Was working on the design of client-server state synchronization and networking for Rookframe. No coding, just reading some nice old articles - this one was very interesting: gamedevs.org/uploads/late...

Threw some test to check whether the foundation holds. Rust + Bevy owns rendering + domain, Bun manages plugin system via a worker pool, communication is TCP for now. 3 plugins move rooks, split them on collisions and use d20 to change direction. Seems to work, but RAM is not good yet =)

Finally got the foundation of my app working. Final count of bugs on verification by GPT-5.6 Sol Ultra - 110. Reasoning for bugs - me. Or rather the way I set up dev process. And the scale of tasks I inputted. So yeah - better task splitting, lots of verification. Reading code =)

Unfortunately even GPT Sol on xhigh and max reasoning still needs babysitting. All development loops had verification and review steps by sub-agents. Still missed 72 bugs.

13h into a GPT-5.6 Sol run at max reasoning: at hour 4, it set itself a goal. In the final hour, it drifted into „fix 1 bug -> heavy review, then fix until reviewer is happy -> full e2e run“. Steering got it back, but checkpoint + fresh context still seems better for long tasks.

First impression of GPT-5.6 Sol/Terra: it allowed me to revive a project that was too complicated and also wouldn’t make much money. But if released, I believe it would be a good thing for the people in one of my hobbies.